Large language models hallucinate. They produce plausible-sounding answers that are factually wrong. This is a feature of how they work, not a bug you can patch. You have to build around it.
In consumer applications, hallucinations are occasionally amusing. In enterprise applications, they are unacceptable. A contract summary that invents a clause. A customer support response that cites a non-existent policy. A financial analysis that misreads a number by an order of magnitude. Any of these can create real damage if they reach the customer or the decision-maker unreviewed.
The answer is not to stop using LLMs. The answer is to scope what they are allowed to say, and to build verification around them.
Scoping means you do not let the LLM answer questions outside the documents you gave it. Retrieval-augmented generation is the technical pattern. You pull the relevant documents, you give them to the model, you tell the model "answer only from these." If the answer is not in the documents, the model is instructed to say it does not know. This cuts hallucination rates by 80 to 90 percent in our experience, and it forces the model to cite its sources.
Verification means that for any answer the model gives, a downstream system checks it. For factual claims, it compares to a ground truth source. For actions, the model proposes the action and a human or a rule engine confirms. For numbers, the model shows its work and the calculation is re-run by code, not by the model itself.
The businesses that get this right treat the LLM as a drafting tool, not an oracle. The LLM drafts. A human or a rule engine verifies. The combination is faster and more accurate than either alone.
The ones that get it wrong treat the LLM as an authoritative answer, deploy it to customers unverified, and discover the hallucination problem in a lawsuit, a regulatory filing, or a viral screenshot. None of those are good days.
Hallucination is real. Hallucination is also manageable. The management pattern is known and not particularly hard. The reason most deployments skip it is time pressure, and the consequences of skipping it show up later, usually at the worst possible time.